Acta Polytechnica Scandinavica, Mathematics, Computing and Management in Engineering Series No. 95, Doctor of Science (Technology) Thesis, Espoo 1998.

Abstract

The goal of the thesis has been to develop a classification system for fast, real-time web surface inspection that learns texture and shape characteristics of target objects from examples without (almost) any a priori information on the target objects. The system has two parts: the segmentation part (defect detection) that is suitable for hardware implementation and the classification part (defect classification) that is made off-line in an operator station.

High speed requirements of a web inspection system limit the number of techniques that can be applied, and the design is often a compromise between the recognition power and the computational complexity. The main contribution of the thesis is the development of fast techniques for defect detection and classification that can satisfy the high speed requirements and that can be applied in a real web inspection system. A new self-organizing map (SOM) variant, the statistical SOM, is proposed for defect detection. It is a fast and simple classifier that can be implemented efficiently in hardware. In addition to a fast classifier, fast feature extraction techniques are also needed. One such group of techniques is called histogram techniques. They have been popular in surface inspection applications as they are relatively simple and fast to compute and yet they have a reasonable performance. Histogram techniques and the SOM in texture analysis, in shape analysis, and in surface inspection are reviewed in the introductory part of the thesis. In the publications these methods have been applied e.g. in paper web inspection.

The developed defect detection scheme offers several improvements over the gray-level thresholding technique that has been traditionally used in commercial web inspection systems. Furthermore, it has been implemented in hardware that is suitable and fast enough to be included in a working web inspection system. The developed defect classification scheme has several features that have not been implemented in web inspection before. The proposed shape and texture features make it possible to do better and more accurate defect classifications. The codebook-based classifier allows addition of new defects into the codebooks continuously. Thus information on different defects can be collected into the classification system over a long period.